1.1 Corpus Creation

1. Scopus download of documents retrieved from search string from Markard et al. (2012), resulting in ca 2.800 documents. Limited to LANGUAGE = ENGLISH AND TYPE = (ARTICLE OR CONFERENCE PAPER OR BOOK OR BOOKCHAPTER OR EDITORIAL OR SHORT SURVEY).
2. Selecting “seed” publications. 1% most cited total + 1% most cited per year + top3 cited papers from each 2015-2017. Ex-post manual insertion and exclusion (authority based). Results in 51 seed papers
3. Retrieving for each seed 1000 publications with most shared references. Again, same limitations as in step 1. Also include tob/1000 cited in the results of step 1.
4. Adittional ex. post filtering based on citations recieved and connectivity in bibliographic coupling network (see below)

That leads to an overal corpus size of:

## Number of unique publications in the final corpus:  12396

1.3 Publications

Note: The following tables refer to the documents in the main corpus. Generally, dgr refers to the degree, n to the number of publications. Subscript .f indicates the number is fractionalized (divided by the number of elements per publication)

1.4 References cited

Note: The following tables refer to the cited references within the corpus. Number of citation always refers to the citations recieved by the documents in the main corpus.

## Number of unique references cited by the final corpus:  26202  (after removing references cited less than 2 times)

This appears to be rather low, only slightly more than 2x more than publications. However, the original number is way higher (ca 200.000), but after excluding (NW based) unconnected or marginally connected references, it drops substantially. This is an indicator that many of the references are cited very few times in the corpus, and the resulting ones represents a somewhat common core.

2.1 Topics by topwords

This might still be finetuned, but initially doesnt look that bad I think. All the topics for me seem to be somewhat identifiable. We should maybe start naming them to make their interpretation later easier.

2.2 Topics over time

Here, we see quite a shift in the discourse over timee. Notably, topic 5 (technology & industry dynamics related) over time lost quite a bit in popularity, while particularly topic 2 (energy & policy), topic 6 (transition management), and topic 9 (policy & governance) gained.

2.3 LDAViz

Here you find a nice way of exploring topics via the LDAVIz tool. It dispolays all topics in a 2 dimensional PCA, and also gives a nice visual representation over the topics top-word distribution and overall frequencies of this words in the corpus. The \(\lambda\) parameter regulates the importance-ordering of the topwords. High \(\lambda\) order words by the highest propability to appear in the topic to the lowest (independent of the overall word popularity in the corpus), whle low \(\lambda\) emphasize words which are very specific to the topic, and rarely appear in others.

Play a bit around. It’s a bit condensed here in the small view, but you can also load it HERE in fullscreen for a better overview.

3.1 Technical description

Note: This analysis refers the co-citation analysis, where the cited references and not the original publications are the unit of analysis. Here, the strength of the relationship between a reference pair \(m\) and \(n\) (\(s_{m,n}^{coc}\)) is expressed by the number of publications \(C\) which are jointly citing reference \(m\) and \(n\).

\[s_{m,n}^{coc} = \sum_i c_{i,m} c_{i,n}\]

The intuition here is that references which are frequently cited together are likely to share commonalities in theory, topic, methodology, or context. It can be interpreted as a measure of similarity as evaluated by other researchers that decide to jointly cite both references. Because the publication process is time-consuming, co-citation is a backward-looking measure, which is appropriate to map the relationship between core literature of a field.

3.2 Identified knowledge Bases

In order to partition networks into components or clusters, we deploy a community detection technique based on the Lovain Algorithm (Blondel et al., 2008). The Lovain Algorithm is a heuristic method that attempts to optimize the modularity of communities within a network by maximizing within- and minimizing between-community connectivity.

We identify the following communities = knowledge bases

It is not the main focus of this exercise, but still informative to see which historical knowledge the fields draws from.

We identify 9 communities of varyying size (note: I just gave it some ad-hoc names to work with for now. Should be revised). Lets take a look at their development over time.

Lets do a first attempt to categorize them. Notice the tabs for details.

4.1 Technical description

This is arguably the most interesting part. Here, we identify the literature’s current knowledge frontier by carrying out a bibliographic coupling analysis of the publications in our corpus. This measure uses bibliographical information of publications to establish a similarity relationship between them. This **coupling-strength} between publications is determined by the number of commonly cited references they share, assuming a common pool of references to indicate similarity in context, methods, or theory. Formally, the strength of the relationship between a publication pair \(i\) and \(j\) (\(s_{i,j}^{bib}\)) is expressed by the number of commonly cited references.

\[ s_{i,j}^{bib} = \sum_m c_{i,m} c_{j,m} \]

Since our corpus contains publications which differ strongly in terms of the number of cited references, we normalize the coupling strength by the Jaccard similarity coefficient. Here, we weight the intercept of two publications’ bibliography (shared refeences) by their union (number of all references cited by either \(i\) or \(j\)). It is bounded between zero and one, where one indicates the two publications to have an identical bibliography, and zero that they do not share any cited reference. Thereby, we prevent publications from having high coupling strength due to a large bibliography (e.g., literature surveys).

\[ S_{i,j}^{jac-bib} =\frac{C(i \cap j)}{C(i \cup j)} = \frac{s_{i,j}^{bib}}{c_i + c_j - s_{i,j}^{bib}} \]

More recent articles have a higher pool of possible references to co-cite to, hence they are more likely to be coupled. Consequently, bibliographic coupling represents a forward looking measure, and the method of choice to identify the current knowledge frontier at the point of analysis.

4.2 Identified research areas

To identify communities in the field’s knowledge frontier (labeled research areas) we again use the Lovain Algorithm (Blondel et al., 2008). We identify the following communities = research areas

4.4 Connectivity between the research areas